JP3354252B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for recognizing an inputted speech by comparing the feature quantity of an inputted speech with the feature quantity of a standard pattern prepared in advance.

[0002]

2. Description of the Related Art In recent years, voice recognition technology for recognizing phonological information of a language generated by humans has been actively studied, and a voice recognition device applying this voice recognition technology to a specific device has been developed. . To recognize phonological information, generally,
Prepare a plurality of standard patterns in units of words or syllables in advance, compare the unknown input voice with each standard pattern, find the standard pattern that is most similar to the input voice,
A method of determining that the standard pattern is a uttered voice is used.

[0003] The technology of recognizing words, syllables, and the like in such a voice recognition technology includes an isolated word voice recognition for recognizing words and the like generated separately and a specific word and the like from continuously generated voices. Recognition can be classified into two types: continuous word speech recognition. When a speech recognition device using a speech recognition technology is put to practical use, it is desirable that the speech recognition device be capable of realizing continuous word speech recognition in consideration of noise generated in the surroundings and unnecessary words that a speaker may generate.

As a continuous word speech recognition technique for recognizing speech by excluding surrounding noise and unnecessary words generated by a speaker, for example, word spotting such as a continuous DP method is conventionally known (Yoshii). Sadahee, "Digital Audio Processing"
Tokai University Press Chapter 8). Here, the word spotting is a technique for finding a unit such as a word or a syllable from a certain voice and extracting a predetermined word. In addition, continuous DP method (continuous Dynam
ic Programming) refers to a method in which an input speech converted into a parameter series such as a spectrum is shifted by one frame from a start end and a standard pattern such as a word or a syllable is DP-matched (Dynamic Programming matchi).
ng), and when the distance as a matching result falls below a certain threshold, continuous word speech recognition in which it is determined that a word or syllable of the standard pattern exists at that time.

Here, when comparing the input voice and the standard pattern in the voice recognition, the input voice is not converted by comparing the voice waveform itself but by removing phase information from the voice waveform and converting it into features related to the spectrum. It is common to handle information. This is because the amount of information is too large if the audio waveform itself is compared, the phase information of the waveform is easily changed by the transmission system and the recording system, and such phase information hardly contributes to human perception of the audio. Because.

As a feature related to a spectrum, a short-time spectrum extracted at regular intervals is generally used. The short-time spectrum means the power spectrum density for each short-time section of voice, and the spectrum envelope is a component that changes slowly according to the frequency, and the spectrum fine structure is a component that changes finely according to the frequency (Sum on a logarithmic scale) for analysis. Among these, the spectral fine structure is unstable under the influence of the pitch and the like. For this reason, in speech recognition, the spectrum envelope is extracted from the short-time spectrum,
It is common practice to make the spectral envelope a feature of speech.

There are various types of techniques for extracting a spectral envelope, and one of them is cepstrum analysis (cepstrum). This cepstrum analysis is defined as the inverse Fourier of the logarithm of the short-time amplitude spectrum of the waveform, and has a feature in that the spectral envelope and the spectral fine structure can be approximately separated. As a technique for extracting a spectral envelope related to cepstrum analysis, in recent years, an attempt has been made to use a cepstrum calculated from a log spectrum resampled at a mel-scale frequency. Such a cepstrum is called a mel cepstrum. Further, as a special type of cepstrum analysis, LPC cepstrum analysis (LPC
Is a linear predictive coding (abbreviation for linear prediction). The LPC cepstrum is a cepstrum calculated directly from a waveform, that is, an FFT cepstrum (FFT is a fast Fourier transfo
rm: an abbreviation for Fast Fourier Transform), which means a cepstrum based on a linear prediction model, and is characterized in that an envelope spectrum in which the spectrum peak is emphasized more than an envelope spectrum based on the FFT cepstrum can be obtained. In other words, focusing on the fact that important information for speech recognition exists at the peak of the spectrum, and emphasizing the peak of the spectrum, the distance scale is made sensitive to achieve more accurate speech recognition. Method.

[0008]

As described above, the detection of the speech feature amount, that is, the extraction of the spectrum envelope of the short-time spectrum uses techniques such as cepstrum analysis, mel cepstrum analysis, and LPC cepstrum analysis. . At this time, a cepstrum coefficient, a mel cepstrum coefficient, and an LPC cepstrum coefficient are used as the audio feature values. However, since such a cepstrum coefficient and the like are feature amounts that do not depend on the input level of the sound, the feature amount of the surrounding noise is reduced to the feature amount of the input sound in a silent section or a section where the sound level is low, such as when a prompt sound is produced. And may cause erroneous recognition. For example, in a silent section, the distance of the word corresponding to the input voice to the standard pattern is widened by the surrounding noise, and the word is mistaken for an uncorresponding word. May be narrowed by the noise and may be mistaken as a corresponding word, which is an obstacle to realizing accurate speech recognition.

[0009]

According to the first aspect of the present invention,
A sound analysis unit that performs acoustic analysis on an input audio signal for each frame to obtain an audio feature value of the audio signal, and audio detection that obtains a degree of audio-likeness for each frame of the input audio signal. Unit, a noise processing unit that converts the acoustic feature of the section determined to have a low degree of voice likeness by the voice detection unit into white noise, a standard pattern storage unit that stores a standard pattern of voice, and a standard pattern storage unit. A matching processing unit is provided for matching the time series of the standard pattern stored in the storage unit with the time series of the acoustic feature amount processed by the noise processing unit.

According to a second aspect of the present invention, in the first aspect of the invention, the sound analysis unit performs a cepstrum analysis for obtaining a cepstrum coefficient as an acoustic feature amount for each frame, and the noise processing unit obtains a cepstrum coefficient. By setting a small cepstrum coefficient for a frame having a low degree of voice likeness, the acoustic feature value is converted to white noise.

According to a third aspect of the present invention, in the first aspect of the present invention, the acoustic analysis unit performs mel-cepstral analysis for obtaining a mel-cepstral coefficient for each frame as an acoustic feature amount, and the noise processing unit performs a mel-cepstral analysis on the speech detecting unit. The acoustic feature is converted to white noise by setting the mel-cepstral coefficient of the obtained frame having a low degree of voice-likeness to be small.

According to a fourth aspect of the present invention, in the first aspect of the present invention, the acoustic analysis unit performs a spectrum analysis for obtaining, for each frame, a short-time spectrum that has been corrected to remove a spectral tilt as an acoustic feature amount. The noise processing unit converts the acoustic feature value into white noise by setting a short-time spectrum of a frame having a low degree of voice likeness obtained by the voice detection unit to be small.

According to a fifth aspect of the present invention, in the first aspect of the invention, the voice detection unit determines that the lower the power of the input acoustic signal is, the lower the level of voice-likeness is, and the noise reduction processing unit determines the level of the voice. As the detection unit determines that the degree of voice-likeness is low, the acoustic feature amount is strongly converted to white noise.

According to a sixth aspect of the present invention, in the first aspect of the invention, the standard pattern stored in the standard pattern storage section is subjected to a process equivalent to white noise conversion of the acoustic feature in the noise processing section. This is the generated standard pattern.

According to a seventh aspect of the present invention, in the first aspect of the invention, the matching processing in the matching processing section is as follows:
This is a word spotting process.

[0016]

According to the first aspect of the present invention, for each frame, the acoustic feature amount of the acoustic signal input by the acoustic analysis unit is obtained, and the degree of soundness of the audio signal is obtained by the audio detection unit. At this time, if there is a section in the acoustic signal where the degree of voice-likeness is determined to be low, the acoustic feature is converted to white noise by the noise processing unit. That is, in the audio signal, a section having a low degree of voice-likeness means a section having no sound or a low sound level. The fact that the acoustic feature is converted to white noise means that the spectrum representing the acoustic feature is smoothed. Therefore, when there is no sound or the sound level is low, the spectrum as the acoustic feature is smoothed, and the influence of the surrounding noise is removed. In the matching processing unit, the time series of the feature amount subjected to such processing is compared with the time series of the standard pattern stored in the standard pattern storage unit, and the matching is performed. Thus, accurate voice recognition is performed regardless of the presence or absence of ambient noise.

According to the second aspect of the present invention, a cepstrum analysis using a cepstrum coefficient as an acoustic feature is selected as an acoustic analysis technique, and white noise of the acoustic feature is realized by reducing the cepstrum coefficient. . According to the third aspect of the present invention, mel-cepstral analysis using a mel-cepstrum coefficient as an acoustic feature is selected as an acoustic analysis technique. By reducing the mel-cepstral coefficient, white noise of the acoustic feature is realized. ing. According to the fourth aspect of the present invention, a spectrum analysis using a spectrum from which a spectrum inclination has been removed as an acoustic feature amount is selected as an acoustic analysis method,
By reducing this spectrum, the acoustic features are converted to white noise. Therefore, according to the second, third, and fourth aspects of the present invention, accurate speech recognition based on stable acoustic features is performed, and the acoustic features are easily converted to white noise.

According to the fifth aspect of the present invention, when determining the degree of voice-likeness in the voice detection unit, it is determined that the lower the power of the input acoustic signal is, the lower the level of voice-likeness is. The degree of white noise conversion of the acoustic feature quantity is determined according to the degree of the power of the acoustic signal that has been performed. In other words, the acoustic signal is converted to white noise as the power is smaller. Thereby, more accurate voice recognition is performed.

According to the present invention, the standard pattern stored in the standard pattern storage section is generated through the same processing as the white noise conversion of the acoustic features in the noise processing section. Is easy to generate. Then, it is possible to prepare a standard pattern that is very similar to the characteristic amount of the acoustic signal actually input, and more accurate speech recognition is performed.

In the seventh aspect of the present invention, the matching unit performs a word spotting process when performing a matching process between the time series of the audio feature quantity and the time series of the standard pattern. Thus, when a word including a word or a syllable generated as a standard pattern is uttered, the word or the like included in the word is extracted and recognized.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram of each unit. An audio input unit 1 for inputting audio has an A / D conversion unit 2 (A / D is ana
logto disital) is connected and this A / D
A sound preprocessing unit 3 and a sound analysis unit 4 are sequentially connected to the conversion unit 2. The sound preprocessing unit 3 is also connected to a sound detection unit 5, and the sound detection unit 5 and the sound analysis unit 4
Is connected to the noise processing unit 6. Then, a standard pattern storage unit 7 is provided, and the standard pattern storage unit 7 and the noise processing unit 6 are connected to a matching processing unit 8, and the matching processing unit 8 is connected to a recognition result output unit 9. .

Here, the audio input unit 1 is, for example, a microphone, and has a structure in which an audio signal, which is an analog signal input from the audio input unit 1, is output to the A / D conversion unit 2.

The A / D converter 2 includes the audio input unit 1
Is converted into a digital signal, and sampling, quantization, and encoding are performed. This A / D
The conditions for digital conversion in the converter 2 are, for example, a sampling frequency of 16 kHz and a quantization bit number of 16 bits. In this case, the sampling is performed after the band limitation of the low frequency band is applied. This is because sampling is performed according to the sampling theorem and aliasing distortion is not performed.
) Is prevented.

Next, the audio pre-processing unit 3 has a structure in which the input audio signal converted into a digital signal by the A / D conversion unit 2 is subjected to high-frequency emphasis (pre-emphasis). The acoustic pre-processing unit 3 is configured by a first-order digital filter having a transfer function of H (z) = 1-Z ~ ¹ ...

Next, the acoustic analysis unit 4 is an arithmetic unit for extracting a characteristic amount of the input audio signal, and has a structure for extracting a spectrum envelope of a short-time spectrum of the audio signal. In this embodiment, a structure is used in which an acoustic signal is subjected to LPC spectrum analysis, and a cepstrum vector (not including the 0th order) ct as a cepstrum coefficient is obtained for each frame. The acoustic signal analysis conditions in the acoustic analyzer 4 are as follows: frame period: 10 ms window period: 16 ms window function: Hamming window LPC analysis order: 14th order Cepstrum order: 14th order

Next, the sound detecting section 5 has a structure for obtaining the degree of soundness of the sound signal based on the average power in each frame of the input sound signal which has been subjected to the high frequency emphasis by the sound preprocessing section 3. It is. The frame average power can be obtained from the zero-order autocorrelation coefficient during the LPC analysis. Here, in the voice detector 5, the relationship between the average frame power p and the degree v of voice likeness determined by the voice detector 5 is defined by the following equation (2).

[0027]

(Equation 2)

P ₀ in this equation 2 is a constant obtained experimentally, and a value slightly larger than the power value at the start and end of the voice section is used. As is clear from this equation,
The voice-likeness degree v is 0 ≦ v ≦ 1, and when the frame average power p is sufficiently large, the voice-likeness degree v is 1, and when the frame average power p is 0, the voice-likeness degree v is 0. , Meanwhile, the frame average power p
Becomes smaller, the degree v of voice-likeness decreases monotonously.

Next, the noise processing section 6 has a structure in which the acoustic feature quantity obtained by the acoustic analysis section 4 is converted into white noise in accordance with the degree of likelihood of voice obtained by the voice detection section 5. is there. In the noise processing section 6, the calculation processing of c * t = vct... Here, as described above, ct
Is a sound feature amount obtained by the sound analysis unit 4, that is, a cepstrum vector, and v is a degree of soundness obtained by the sound detection unit 5. And
c * t is a cepstrum vector converted into white noise according to the degree of likelihood of speech in the input acoustic signal. As is clear from this equation, the noise processing unit 6 determines the cepstrum vector c * t by the product of the cepstrum vector ct and the degree v of voice-likeness. Here, the cepstrum vector of the white noise is 0, that is, 0
Vector. Therefore, the lower the voice-likeness degree v, the stronger the cepstrum vector ct becomes white noise.

Next, the standard pattern storage section 7 stores
A large number of standard patterns, such as words and syllables, for which speech recognition is executed are stored. These standard patterns are input to the voice input unit 1 and converted into cepstrum vectors ct by the acoustic analysis unit 4, and have the same contents as the time series of the cepstrum vectors c * t subjected to the predetermined processing by the noise processing unit 4. Has,
This is a cepstrum vector c * r generated through the same processing as the time series of the cepstrum vector c * t.

Next, the matching processing unit 8 calculates the time series of the standard pattern stored in the standard pattern storage unit 7, that is, the cepstrum vector c * r, the acoustic feature amount processed by the noise processing unit 6, That is, it has a structure in which matching processing is performed with the time series of the cepstrum vector c * t. The matching process in the matching processing unit 8 is a matching process using the continuous DP method. At this time, a distance scale such as a group delay spectrum distance scale is used as the distance scale.

Next, the recognition result output unit 9 has a structure for outputting the recognition result of the matching processing unit 8,
For example, the structure is such that the presence or absence of a corresponding word or the like is output as a signal or display.

In such a configuration, the audio signal input to the audio input unit 1 is digitally converted by the A / D converter 2 and is sampled, quantized, and coded. Then, high-frequency emphasis is performed by the acoustic preprocessing unit 3, and the spectrum inclination is flattened. As a result, the dynamic range of the audio signal is compressed, and the effective SNR (signal-to-quantization noise) is reduced.
ise ratio (signal to quantization noise ratio) is increased.

Next, the characteristic amount of the high-frequency emphasized sound signal is extracted as a cepstrum vector ct by the LPC cepstrum analysis by the sound analysis unit 4. At the same time, the voice detection unit 5 calculates the voice-like degree v of the high-frequency-emphasized audio signal based on the average power p of each frame by Expression 2.

The cepstrum vector ct and the degree v of voice-likeness thus determined are converted to the noise processing unit 6.
The cepstrum vector c * t which has been subjected to white noise processing by the arithmetic processing of Expression 3 in the noise processing section 6 is obtained. Here, the cepstrum vector c * t processed by the noise reduction processing unit 6 is converted to white noise more strongly as the degree of voice likeness obtained by the voice detection unit 5 is lower. In other words, low audioness means
Since the section is silent or has a low voice level, the silent section or the section having a low voice level is converted to white noise, and the spectrum of the section is flattened.

Next, in the matching processing unit 8, the time series of the cepstrum vector c * t that has been processed by the noise processing unit 6 and the cepstrum vector c * r that is the standard pattern stored in the standard pattern storage unit 7 Is subjected to a matching process. The matching process at this time is performed by the continuous DP method that is word spotting. Therefore, speech end point-free speech recognition is performed. Then, the cepstrum vector c * t to be matched
Both the time series of the cepstrum vector c * r and the time series of the cepstrum vector c * r have a low degree of voice-likeness, that is, a section with no sound or a low voice level is converted to white noise, and the spectrum of the section is smoothed. ing. Therefore, voice recognition is performed without being affected by ambient noise, and the accuracy of voice recognition is improved. Therefore, the matching processing unit 8
A highly accurate recognition result is output from the recognition result output unit 9 that outputs the processing result of (1).

Here, a modified example of the acoustic analysis unit 4 will be described. In the present embodiment, the acoustic analysis unit 4 for performing the LPC cepstrum analysis is provided as a method for obtaining the characteristic amount of the input audio signal. However, the method for obtaining the characteristic amount of the audio signal is not limited thereto. An acoustic analysis unit using a technique such as analysis, mel-cepstral analysis, or spectral analysis with spectral tilt correction may be used.
In short, any type can be used as long as it is a structure that can obtain a feature amount that can facilitate the white noise processing in the noise processing unit 6. More specifically, in the mel cepstrum vector as the mel cepstrum coefficient, the 0 vector represents white noise, similarly to the cepstrum vector. The spectrum analysis with the spectrum tilt correction is performed by performing a spectrum tilt correction such as a logarithmic conversion or a correction (exponential conversion) for reducing a least square approximation straight line on the spectrum obtained by the FFT or the band-pass filter bank. Be executed. As a result, in the corrected spectrum vector, the 0 vector expresses white noise as in the case of the cepstrum vector. Therefore, the standard pattern stored in the standard pattern storage unit can be directly applied to the apparatus of the present embodiment simply by using a standard pattern subjected to mel-cepstral analysis or a standard pattern subjected to spectral analysis with spectral tilt correction. .

Next, a modified example of the voice detection unit 5 will be described. First, the voice detection unit 5 calculates the average power P of each frame as basic data for obtaining the degree of voice-likeness in the audio signal.
Is obtained by the LPC analysis, the voice detection unit 5 may be configured by partially sharing the structure of the acoustic analysis unit 4 that performs the LPC cepstrum analysis. As another modified example of the voice detection unit 5, the voice power is not used as basic data for determining the degree of voice-likeness, but the number of zero crossings, pitch frequency, sharpness of formants, distance from each phoneme pattern, and the like are used. Is also good.

Next, a modification of the matching processing section 8 will be described. In the present embodiment, the matching processing unit 8 is configured to execute the continuous DP method. However, the matching processing unit 8 may be configured to execute matching by another method using a state transition model or the like, and is not limited to such word spotting. Instead, a structure for executing isolated word speech recognition may be adopted.

[0040]

According to the first aspect of the present invention, there is provided an acoustic analysis unit for performing acoustic analysis on an input audio signal for each frame to obtain an acoustic feature of the audio signal, and A voice detection unit that calculates the degree of voice-likeness for each frame, a noise processing unit that converts the acoustic features of the section determined to have low voice-likeness into white noise by the voice detection unit, and a voice standard. A standard pattern storage unit for storing a pattern, and a matching processing unit for matching a time series of the standard pattern stored in the standard pattern storage unit with a time series of the acoustic feature amount processed by the noise reduction processing unit. Therefore, if there is a section in the input sound signal where the degree of voice-likeness is determined to be low, that is, if there is a section where there is no sound or the sound level is low, the audio feature The processing unit smoothes the spectrum that expresses the acoustic features by converting it to white noise, thereby removing the influence of surrounding noise on that section, and preventing erroneous recognition due to the surrounding noise being mixed with the acoustic features. Therefore, the present invention has an effect that the accuracy of voice recognition can be improved.

According to a second aspect of the present invention, in the first aspect of the present invention, the sound analysis unit performs a cepstrum analysis for obtaining a cepstrum coefficient as an acoustic feature amount for each frame, and the noise processing unit obtains a cepstrum coefficient. By setting the cepstrum coefficient of a frame having a low voice-likeness to a small value, the acoustic feature is converted to white noise. According to the third aspect of the present invention, in the first aspect of the present invention, Perform a mel-cepstral analysis to determine the cepstrum coefficient for each frame, and in the noise reduction processing unit, set the mel-cepstral coefficient of the frame with a low degree of voice-likeness obtained by the voice detection unit to a small value to make the acoustic feature amount white noise, According to a fourth aspect of the present invention, in the first aspect of the present invention, the acoustic analysis unit includes a spectrum gradient as an acoustic feature. The short-time spectrum correction is performed to remove performs a spectrum analysis for determining for each frame,
The noise reduction processing unit is configured to convert the acoustic features into white noise by setting the short-time spectrum of the frames with low voice-likeness determined by the speech detection unit to be small, so that the noise based on the stable acoustic features Accurate speech recognition can be performed, so that the speech recognition accuracy can be further improved, and the acoustic feature can be easily converted to white noise. Can be simplified.

According to a fifth aspect of the present invention, in the first aspect of the present invention, the voice detection unit determines that the lower the power of the input audio signal is, the lower the level of voice-likeness is. The acoustic feature is configured to strongly white noise as the detection unit determines that the degree of voice-likeness is low, so it is easy and accurate to use the power of the acoustic signal as a parameter to determine the degree of voice-likeness. The determination of the degree of voice-likeness can be realized, and therefore, the arithmetic processing can be simplified. In addition, the smaller the power of the sound signal is, the stronger the white noise becomes, and thus the more accurate the sound signal is. It has effects such as being able to contribute to the realization of recognition.

According to a sixth aspect of the present invention, in the first aspect of the present invention, the standard pattern stored in the standard pattern storage section is subjected to processing equivalent to white noise conversion of the acoustic feature in the noise processing section. Since the generated standard pattern is used, it is easy to generate the standard pattern, and it is possible to prepare a standard pattern that is extremely similar to the characteristic amount of the actually input audio signal, and therefore, a more accurate voice It has effects such as being able to contribute to the realization of recognition.

According to a seventh aspect of the present invention, in the first aspect of the present invention, the matching processing in the matching processing section is as follows:
Since the word spotting process is performed, when a word including a word or a syllable generated as a standard pattern is uttered, the word or the like included in the word can be extracted and recognized. This has an effect that it is possible to prevent an erroneous recognition result from occurring when there is no input voice.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 4 Acoustic analysis unit 5 Voice detection unit 6 Noise reduction processing unit 7 Standard pattern storage unit 8 Matching processing unit

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 15/10 G10L 21/02 G10L 15/20

Claims (7)

(57) [Claims] An acoustic analysis unit for performing acoustic analysis on an input audio signal for each frame to obtain an acoustic feature amount of the audio signal, and an audio-likeness of the input audio signal for each frame. A voice detection unit for determining the degree, a noise processing unit for converting the acoustic feature amount of a section determined to have a low degree of voice likeness by the voice detection unit into white noise, and a standard pattern storage unit for storing a standard pattern of voice And a matching processing unit for matching the time series of the standard pattern stored in the standard pattern storage unit with the time series of the acoustic feature amount processed by the noise processing unit. apparatus. 2. A sound analysis unit performs a cepstrum analysis for obtaining a cepstrum coefficient as an acoustic feature amount for each frame, and a noise reduction processing unit sets a small cepstrum coefficient of a frame having a low degree of soundness obtained by the sound detection unit. 2. The speech recognition apparatus according to claim 1, wherein the acoustic feature is converted into white noise by performing the processing. 3. A sound analysis unit performs a mel cepstrum analysis for obtaining a mel cepstrum coefficient as an acoustic feature amount for each frame, and a noise reduction processing unit performs a mel cepstrum coefficient calculation for a frame having a low degree of voice likeness obtained by the sound detection unit. 2. The speech recognition apparatus according to claim 1, wherein the acoustic feature value is converted to white noise by setting a small value. 4. A sound analysis unit performs a spectrum analysis for each frame to obtain a short-time spectrum corrected for removing a spectrum inclination as an acoustic feature amount, and a noise processing unit performs a speech likeness obtained by a speech detection unit. 2. The speech recognition apparatus according to claim 1, wherein the acoustic feature quantity is converted into white noise by setting a short-time spectrum of a frame having a low degree of noise to be small. 5. The voice detection unit determines that the lower the power of the input audio signal is, the lower the level of voice-likeness is, and the noise processing unit determines that the level of voice-likeness is lower by the voice detection unit. The speech recognition apparatus according to claim 1, wherein the acoustic feature amount is strongly converted to white noise. 6. The standard pattern stored in the standard pattern storage unit is a standard pattern generated through processing equivalent to white noise conversion of an acoustic feature in a noise processing unit. The speech recognition device according to claim 1. 7. The speech recognition device according to claim 1, wherein the matching processing in the matching processing unit is word spotting processing.